genderBR

This is a new version that introduces a new feature: asymmetrical thresholds for gender classification. In the previous versions, users could set a single threshold to classify names as female or male -- and they could overlap if the threshold was too low. Now, users can set a different threshold for each sex, and there is a new check to prevent overlapping thresholds. The release also includes other improvements and bug fixes.

• The threshold argument of get_gender() and get_gender_nn() now accepts a vector with two values, allowing users to set a different threshold for each sex: the first value is used to classify females and the second, males (e.g., threshold = c(0.9, 0.8)). The two values can also be named, in any order (c(Female = 0.9, Male = 0.8) or c(F = 0.9, M = 0.8)). A single value keeps setting a symmetrical threshold, as before.
• Thresholds that sum to less than 1 are now rejected with an informative error. These values make the female and male bands overlap, so a name could be classified as both. This also applies to a single threshold below 0.5, which was silently accepted before and gave inconsistent results between get_gender() and get_gender_nn().
• Added an nn_size argument to get_gender_nn() and get_gender(nn = TRUE) that splits a large input vector of names into batches. This keeps avoids out-of-memory crashes on big vectors. Defaults to NULL (all names classified in a single pass).
• Added a device argument to get_gender_nn() and get_gender(nn = TRUE) to run neural network inference on a GPU ("cuda" or "mps") instead of the default CPU for faster predictions.
• Fixed a vocabulary mismatch in the neural network encoder: characters not in the model vocabulary were mapped to <PAD> instead of the <UNK>. Names containing rare characters are now handled correctly.
• Moved torch from Imports to Suggests. The neural network model is now built only when the NN backend (get_gender_nn() or get_gender(nn = TRUE)) is used. A clear message tells users to install torch when it is not available. The default methods no longer require torch.
• Added new tests for the new features and fixes.